부산대학교 도서관

3D reconstruction is the key technology to emerging technologies such as smart robotics, VR/AR/XR and autonomous driving. To enhance the robustness of our proposed 3D reconstruction system, the HDR-based SLAM is adopted in the camera pose estimation step to improve the qualitative result of geometric reconstruction. The proposed HDR-based SLAM uses the pre-calibrated inverse camera response function (CRF) to map a single RGB image into a radiance map. To exclude the influence of exposure time, normalized radiance maps independent of exposure time are used during tracking. Since ORB feature matching is the basic element of tracking and mapping in our system, the ORB descriptor patch is re-trained especially for normalized radiance maps. Experimental results have shown good performance of our system under challenging low-light environment, which helps expand the applicability of 3D reconstruction system.